https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112900
Bug ID: 112900
Summary: Missing optimization: canonicalize `select c, x - 1, x
+ 1` to `x + (select c, -1, 1)` (or reversely)
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: xxs_chy at outlook dot com
Target Milestone: ---
GCC works much better on folding branches into select than llvm for:
https://godbolt.org/z/jWzePjqTs
But GCC seems to generate different X86 assembly for the code
below(https://godbolt.org/z/Erq56Tbjq):
int src(int x, int y, bool cond) {
return (x > y ? x - 1 : x + 1);
}
int tgt(int x, int y, bool cond) {
return x + (x > y ? -1 : 1);
}
For `src`, we compute both `x-1` and `x+1` and apply `cmov` to return one of
them.
For `tgt`, we fold `(x > y ? -1 : 1)` to `(x > y) * 2 - 1`.
LLVM prefers the latter, so I think it may be better to canonicalize `select c,
x - 1, x + 1` to `x + (select c, -1, 1)`. Though it's uncanonicalized in
tree-optimization stage, please let me know if there is other backend's factors
like ILP deciding that.
